Skip to main page content
Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
, 78 (6), 2275-2282

Feasibility of Through-Time Spiral Generalized Autocalibrating Partial Parallel Acquisition for Low Latency Accelerated Real-Time MRI of Speech

Affiliations

Feasibility of Through-Time Spiral Generalized Autocalibrating Partial Parallel Acquisition for Low Latency Accelerated Real-Time MRI of Speech

Sajan Goud Lingala et al. Magn Reson Med.

Abstract

Purpose: To evaluate the feasibility of through-time spiral generalized autocalibrating partial parallel acquisition (GRAPPA) for low-latency accelerated real-time MRI of speech.

Methods: Through-time spiral GRAPPA (spiral GRAPPA), a fast linear reconstruction method, is applied to spiral (k-t) data acquired from an eight-channel custom upper-airway coil. Fully sampled data were retrospectively down-sampled to evaluate spiral GRAPPA at undersampling factors R = 2 to 6. Pseudo-golden-angle spiral acquisitions were used for prospective studies. Three subjects were imaged while performing a range of speech tasks that involved rapid articulator movements, including fluent speech and beat-boxing. Spiral GRAPPA was compared with view sharing, and a parallel imaging and compressed sensing (PI-CS) method.

Results: Spiral GRAPPA captured spatiotemporal dynamics of vocal tract articulators at undersampling factors ≤4. Spiral GRAPPA at 18 ms/frame and 2.4 mm2 /pixel outperformed view sharing in depicting rapidly moving articulators. Spiral GRAPPA and PI-CS provided equivalent temporal fidelity. Reconstruction latency per frame was 14 ms for view sharing and 116 ms for spiral GRAPPA, using a single processor. Spiral GRAPPA kept up with the MRI data rate of 18ms/frame with eight processors. PI-CS required 17 minutes to reconstruct 5 seconds of dynamic data.

Conclusion: Spiral GRAPPA enabled 4-fold accelerated real-time MRI of speech with a low reconstruction latency. This approach is applicable to wide range of speech RT-MRI experiments that benefit from real-time feedback while visualizing rapid articulator movement. Magn Reson Med 78:2275-2282, 2017. © 2017 International Society for Magnetic Resonance in Medicine.

Keywords: low latency imaging; real-time MRI; speech production.

Figures

FIG. 1.
FIG. 1.
k-t spiral sampling in prospective studies: pseudo-golden-angle sampling was considered with angle increments of 222.490 and a period of 13 interleaves. In the calibration stage (a), the interleaves are sorted in ascending order of the distribution. In the reconstruction stage (b), the interleaves are acquired with golden-angle increments.
FIG. 2.
FIG. 2.
Effect of calibration size on spiral GRAPPA reconstruction: (a) nRMSE; (b) SSIM between the reconstructions and fully sampled images is plotted v/s the number of time frames during calibration (tcalib). (c) Example spatial frame from reconstructions using fully sampled data, zero filled at R = 4, spiral GRAPPA at R = 4 with different tcalib. nRMSE and SSIM respectively decreased and increased monotonically with increasing tcalib, however plateaus at tcalib = 150. Similar reconstructions were observed with tcalib > 150 (see blue highlighted area that corresponds to nMRSE within 0.3% and SSIM within 0.9%). In this work, we use tcalib = 150, which is ~11.7 seconds.
FIG. 3.
FIG. 3.
Spiral GRAPPA reconstructions from retrospectively undersampled spiral data (12 interleaves/frame, 2.5 mm2). The speaker repeated the phrase “ala-ara-asa-asha” at a normal pace. The first column shows an example spatial frame; the second column shows the image time profile marked by the solid white arrow; the third column shows the error time profiles, which are scaled up ×3 for better visualization. Good spatiotemporal fidelity is maintained up to R = 4. At R > = 6, spatial blurring of the high-frequency edges is visually evident. The noise amplification with increasing R was consistent with the factor Nfs/Nus; where Nfs and Nus are respectively the number of acquired interleaves in the fully sampled data and undersampled data.
FIG. 4.
FIG. 4.
RT-MRI of fluent speech with (a) zero-filled reconstruction at R = 4.3 (3 TR, 18 ms); (b) view-sharing reconstruction (step size: 3 TR; frame rate: 55 fps; native time resolution = 78 ms); (c) through-time GRAPPA reconstruction at R = 4.3 (3 TR, 18 ms). The subject produced the phrase “one-two-three-four-five” at a rapid pace. The first three columns show a sequence of spatial frames corresponding to the timing of the tongue tip hitting the palate, and the last column shows the image time profiles at the cross-section marked by the white dotted arrow. Zero-filled–based images show considerable undersampling artifacts. View sharing captured slow dynamic movements well, but blurred rapid movements. In contrast, through-time spiral GRAPPA provided improved temporal fidelity of rapidly moving articulators. For instance, subtle tongue tip shaping is seen to be captured with spiral GRAPPA while blurred with view sharing (see solid yellow arrows).
FIG. 5.
FIG. 5.
RT-MRI of beat boxing at 55 fps with (a) view sharing, (b) spiral GRAPPA, and (c) PI-CS. The dynamic time series corresponded to the 5-second snippet of a free-style beat-boxing task. The last column shows the image time profile as marked by the solid white arrow. Columns 1 to 4 show the dynamic images of the time instance marked on the first column. In comparison to PI-CS, and spiral GRAPPA, view sharing resulted in temporal blurring of fast articulatory dynamics. For instance, the rapid inward movement of the tongue tip is captured with better fidelity in the spiral GRAPPA and PI-CS methods when compared to view sharing (see white arrows). Similarly, subtle movements of the velum are blurred with view sharing in comparison to spiral GRAPPA and PI-CS (see yellow arrows). PI-CS has improved signal to noise attributed to inherent denoising. Approximate MATLAB (The MathWorks, Inc., Natick, MA, USA) latency times were: view sharing (14 ms/frame); spiral GRAPPA (116 ms/frame); and PI-CS (17.5 mins to jointly reconstruct 5 secs of RT-MRI data).

Similar articles

See all similar articles

Cited by 4 articles

Feedback